A new arXiv paper, titled "FLINT: Fixed-budget Late Interaction," addresses a known trade-off in multi-vector retrieval models. Systems such as ColBERT match queries and documents at the token level, which often gives them an accuracy advantage over single-vector models of similar size.
The abstract notes that this gain comes at a price: retaining a vector for every token substantially increases storage and computational demands. FLINT appears to tackle this by imposing a fixed budget on the number of vectors used per document, aiming to preserve the benefits of late interaction while controlling costs.
The source is limited to the abstract, so the specific mechanism and experimental results are not detailed here. The paper is listed as a new arXiv announcement under ID 2610.27273.